Graph-based deep reinforcement learning for haplotype assembly with Ralphi.

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Machine Learning in Biology

Background:

• Haplotype assembly is crucial for understanding genotype-phenotype correlations.
• Existing methods struggle with accuracy and scalability for individual diploid genomes.
• Read-based approaches reconstruct haplotypes directly from sequencing data.

Purpose of the Study:

• To introduce Ralphi, a novel deep reinforcement learning framework for read-based haplotype assembly.
• To improve the accuracy of reconstructing maternal and paternal inherited chromosome copies.
• To enable a better understanding of how allele combinations influence biological traits.

Main Methods:

• Developed Ralphi, integrating deep learning and reinforcement learning for read fragment partitioning.
• Utilized the maximum fragment cut formulation on fragment graphs for RL reward objectives.
• Trained Ralphi on diverse fragment graph topologies from the 1000 Genomes Project.

Main Results:

• Ralphi demonstrated lower error rates compared to state-of-the-art methods.
• Achieved comparable or longer haplotype block lengths across various coverage levels.
• Validated performance on standard human genome benchmarks for both short and long reads.

Conclusions:

• Ralphi offers a powerful new approach for accurate haplotype assembly.
• The framework shows significant improvements over existing methods for individual diploid genomes.
• This advancement facilitates deeper insights into genetic variation and its phenotypic impact.